[bedrock] Add Claude Opus/Sonnet 5.5 to mantle, fix 5.5 inference profiles, add gpt-6.1-sol to OpenRouter

[naresh4dev]

Summary

Adds Claude Opus 5.5 / Sonnet 5.5 to bedrock-mantle, corrects the Bedrock inference profiles and params for the Claude 5.5 family, adds gpt-6.1-sol to OpenRouter, and normalizes nine prices that were written in scientific notation.

Every price below was taken from an official source and converted with dollars_per_million / 10000.

1. bedrock-mantle: add claude-opus-5-5 and claude-sonnet-5-5

Both model cards list bedrock-mantle support (opus 5.5, sonnet 5.5).

Model	Input	Output	5m cache write	1h cache write	Cache read
anthropic.claude-opus-5-5	$4 → 0.0004	$20 → 0.002	$5 → 0.0005	$8 → 0.0008	$0.20 → 0.00002
anthropic.claude-sonnet-5-5	$2 → 0.0002	$10 → 0.001	$2.50 → 0.00025	$4 → 0.0004	$0.20 → 0.00002

No batch_config — mantle doesn't expose the Batch API, consistent with all 59 pre-existing entries in that file.

2. bedrock: remove Sonnet 5.5 profiles that don't exist

The model card lists Sonnet 5.5 on bedrock-runtime as Global-CRIS-only:

Endpoint	Model ID	In-Region	Geo	Global
bedrock-runtime	anthropic.claude-sonnet-5-5	N/A	N/A	global.anthropic.claude-sonnet-5-5

f81ffd1f added bare / us. / eu. variants AWS doesn't offer, so calls against them fail upstream while still resolving a local price. The us./eu. entries also carried a +10% markup that a global-only listing can't justify. Removed from general/ and pricing/ (incl. -v1 aliases); global. retained at the official rates. Also dropped its batch_config — the model card marks Batch unsupported.

This applies the same rule as c158d7da (bare openai.gpt-6.1-sol dropped, in-region unsupported) and bb6853bf (global dropped, unsupported).

Cost attribution is unaffected for any prefixed id — the gateway already falls back to the region-stripped key:

// gateway-enterprise-node/src/services/winky/handlers/modelConfig.ts
if (!entry && provider === BEDROCK) {
  const strippedModel = getBedrockModelWithoutRegion(model);
  if (strippedModel !== model) entry = pricingConfig?.[strippedModel];
}

3. bedrock: fix Opus 5.5 params + add au./jp.

Opus 5.5 has adaptive thinking always on; thinking.budget_tokens and thinking:{type:"disabled"} both return 400, and forced tool choice errors (thinking support matrix). All four entries still used the Opus 5 enabled/disabled + budget_tokens shape. Switched to adaptive + output_config.effort (default medium) and trimmed tool_choice to none/auto — matching the existing anthropic.claude-fable-5-1 entry in the same file.

Added the au. and jp. geo profiles from the model card, priced like the existing us./eu. entries at $4.40/$22 (regional carries a 10% premium over global, per Anthropic's Bedrock docs and models.dev).

Sonnet 5.5 gets adaptive | between_tools instead — it's the only model that accepts between_tools, in place of disabled.

4. openrouter: add gpt-6.1-sol

Served via the OpenAI and Azure upstreams (source): $2/$10 in/out, $2.50 cache write, $0.10 cache read → 0.0002 / 0.001 / 0.00025 / 0.00001.

This was the last confirmed gap — openai, open-ai, azure-openai, azure-ai, bedrock (us. only) and bedrock-mantle already had it. I could not confirm 6.1 on Lightning AI's hub, so it was left alone rather than guessed.

5. Scientific notation → plain decimals

Nine literals were in exponent form, which the pricing consumer doesn't parse:

File	Value	Fixed
pricing/x-ai.json — grok-4.7 cache read (×2)	5e-5	0.00005
pricing/workers-ai.json — bge-reranker-base	3e-7	0.0000003
pricing/together-ai.json — m2-bert retrieval (×6)	8e-7	0.0000008

Numeric values unchanged — verified by diffing parsed JSON before/after. Formatting only.

Testing

• jq empty on all 9 touched files — valid
• python3 scripts/check_duplicate_keys.py — 112 files, no duplicates
• npm run format && npm run lint — no new errors (the 2653 in general/scenario-ai.json are pre-existing on main; I reverted prettier's unrelated reformat of that file to keep this diff focused)
• Programmatic key-level diff vs main confirms exactly 6 additions, 9 removals and 7 intentional modifications, with no unintended entries touched
• rg confirms zero scientific-notation prices remain repo-wide

Notes for reviewers

While auditing I found the same inflated us.-prefix pattern on openai.gpt-6-sol, xai.grok-4.6/4.7 and moonshotai.kimi-k3, where the us. entry is +10% over a base that AWS docs say shouldn't carry a premium. Left untouched — flagging for a separate pass.

[Copilot]

Copilot review overview

🔵 Needs a closer look

The model availability, parameter contracts, and billing rates depend on current provider documentation that requires final human verification.

Review effort: Balanced
Findings: None

What changed in this PR

Adds new Bedrock Mantle and OpenRouter models while correcting Claude 5.5 Bedrock profiles, parameters, and pricing representation.

Changes:

• Adds Claude Opus/Sonnet 5.5 to Bedrock Mantle and GPT-6.1 Sol to OpenRouter.
• Corrects Bedrock Claude 5.5 profiles, pricing, batch support, and inference parameters.
• Replaces scientific-notation prices with plain decimals.

File	Description
general/​bedrock.json	Updates Claude 5.5 profiles and parameters.
general/​bedrock-mantle.json	Adds Claude Opus/Sonnet 5.5 metadata.
general/​openrouter.json	Adds GPT-6.1 Sol metadata.
pricing/​bedrock.json	Corrects Claude 5.5 profiles and pricing.
pricing/​bedrock-mantle.json	Adds Claude 5.5 pricing.
pricing/​openrouter.json	Adds GPT-6.1 Sol pricing.
pricing/​x-ai.json	Normalizes decimal notation.
pricing/​workers-ai.json	Normalizes decimal notation.
pricing/​together-ai.json	Normalizes decimal notation.

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.